Garrett's personal free-GPU studio suite, built on Hugging Face's free tier.
Personal tooling only. Nothing here is connected to, or part of, any
Galaxy Sports Edge product, database, Stripe, or customer data — and
it never will be.
Live & verified
Studio Chat
General chat with presets for code, writing, and structured JSON output.
Model: Qwen2.5-72B-Instruct, 4-bit (nf4) · License: Qwen (custom)
Packed size: 44.5GB
Measured: ~13.4 tok/s, 200-token Code-preset reply in 40.8s
Limits: max 1024 output tokens/reply, 120s ZeroGPU cap
Open studio-chat →
Live & verified
Studio Media
Image generation, speech transcription, text-to-speech, and voice cloning.
Image: FLUX.1-dev (non-commercial license, personal use only) — 1024x1024 in 22.6–27.8s, batch of 4 in 102.8s; optional AntiBlur detail-boost LoRA and Detail Daemon sigma-schedule texture boost, both off by default and both real-call verified with visible before/after; prompt enhancer via studio-chat in 17.9s
Upscale: Real-ESRGAN 2x/4x (BSD-3-Clause) — 2x in 9.4s; optional GFPGAN face restore (for degraded photos only — regresses already-sharp FLUX output, see README)
Remove Background: BiRefNet (MIT) — transparent-PNG cutout in 7.5s, clean edges verified on a real generated image
Transcribe: Whisper-large-v3-turbo (MIT) — SRT/VTT + word timestamps in 7.9s
Speak: Kokoro-82M (Apache-2.0) — in 8.8s
Clone: F5-TTS (CC-BY-NC-4.0, personal use) — own-voice only, consent enforced server-side, watermarked filename, in 13.1s
Music: MusicGen-small (CC-BY-NC-4.0, personal use) — 10s clip in 12.5s, real audio verified
Packed size: ~36GB (Vision tab, FLUX.1-schnell "fast draft", and Chatterbox-Turbo TTS all tried and reverted — packing budget, VRAM capacity, and a hard diffusers version conflict respectively; see the Space's README)
Open studio-media →
Live & verified
GSE Proof MCP
Independent, read-only MCP server over Galaxy Sports Edge's public proof/verification endpoints.
Deployed on HF PRO (cpu-basic) after the free-tier 402 blocker was resolved
5 MCP tools, all real-call verified: ledger, receipts list, receipt verify, calibration status, board coverage
No auth, no writes, rate-limited, 60s cache
Open gse-proof-mcp →
Deployed
Podcast Pipeline
Markdown/PDF/DOCX script + your voice clip → narrated episode. Plus Clip Studio: cut, reframe, and caption your own video.
Deployed on HF PRO (cpu-basic), calls studio-chat and studio-media — no local models
Clip Studio, real-call verified: scene-detect (PySceneDetect, BSD-3) in 7.5s, 9:16/1:1/16:9 reframe in 2.4s, caption burn (via studio-media Whisper) in 12.1s
Full episode production: not yet run end-to-end at real script length
Open podcast-pipeline →
Code-complete, hardware pending
Studio Video
Text-to-video generation on ZeroGPU.
Model: LTX-Video-0.9.7-distilled (Lightricks, LTX-Video Open Weights License, personal use) — diffusers-native, 7 recommended inference steps
Blocked: this session's OAuth token can create/upload to the Space but the Hub's hardware-change endpoint (zero-a10g) 401s for OAuth tokens — a real platform restriction, not a code bug (confirmed: the Space's own code loads and fails cleanly with "no NVIDIA driver" on the cpu-basic fallback, proving the code path itself is sound)
One manual step needed: Space Settings → change hardware to ZeroGPU
Open studio-video →